Research On Breast Cancer Pathological Image Classification Based On Deep Learning Method

Breast cancer is a serious threat to women’s health,and early diagnosis and treatment can improve patients’ survival rates.Common diagnostic methods include imaging analysis and histopathological analysis of tissue biopsy.The former includes magnetic resonance imaging and mammography X-ray examination,among other methods.Compared to the former,histopathological analysis of tissue biopsy is considered the "gold standard" for diagnosing breast cancer,as it can provide authoritative diagnostic evidence for doctors.Traditional pathological diagnosis relies on manual review of images at different magnifications,which has some subjectivity and variability.In recent years,the rapid development of artificial intelligence has provided new auxiliary approaches for pathological diagnosis.This method can help doctors improve diagnostic accuracy and make diagnostic results more scientific and objective.Among them,methods represented by deep learning have made good progress in the field of benign and malignant classification of pathological images,but there are still some problems,the most typical of which include:1)Manually annotating breast pathology images is time-consuming and labor-intensive,and small sample datasets are common in reality,which can lead to neural networks being prone to underfitting and having low generalization ability;2)Spatial information of breast tissue in pathological images is one of the important diagnostic criteria for pathologists to diagnose breast cancer,but convolutional neural networks have difficulty fully perceiving spatial information;3)Some benign and malignant tumors have similar appearances in pathological images,with small feature differences.Traditional deep neural networks have insufficient discriminative ability for benign and malignant pathological images with small feature differences.This paper focuses on the research of benign and malignant classification in breast pathology images using deep learning methods.The specific contents include:(1)Research on Benign and Malignant Classification of Pathological Images Based on MultiStage Transfer Learning and Attention Mechanism: This paper adopts VGG19 and Res Net34 as the basis and carries out research using a multi-stage transfer learning strategy in two stages to address the first problem mentioned.In the first stage of transfer learning,based on the characteristics of sharing low-level information such as texture features between natural images and pathological images,a model with good pathological feature extraction performance is trained to classify a largescale IDC breast pathology image dataset by introducing Image Net pre-training parameters.After the above training,in the second stage of transfer learning,the model is further fine-tuned to improve its classification performance on small-sample datasets.The accuracy reached 0.895 on the Breakhis dataset.Attention mechanism is introduced based on VGG19 and Res Net34 to select important information from the channels and spatial levels of the image,thereby improving the model’s classification performance.The model achieved an accuracy of 0.888 in the pathological image classification task.By combining the above two methods,the optimal accuracy of the model was further improved to 0.914.(2)Research on Benign and Malignant Pathological Image Classification Based on Graph Convolutional Network: In response to problem 2 mentioned above,this paper conducts research based on graph convolutional networks.The commonly used methods of constructing graphs are complex.so this paper uses the feature map obtained by convolutional neural networks to construct a graph structure.This structure can fully reflect the spatial correlation of different regions of pathological tissues,and can improve the model’s perception ability of spatial feature information.Ordinary graph convolutional networks have a limited number of convolutional layers,and their performance in extracting deep-level features is weak.Therefore,using the idea of alternate updating mechanism,the convolutional layers are formed into a cyclic feedback structure,and any two convolutional layers can perform forward propagation and backward connection.Feature information is transmitted between any layers,and through repeated superposition,the network extracts deeperlevel features.The result indicates that the classification accuracy of the model has reached 0.919.(3)Research on Benign and Malignant Pathological Image Classification Based on Contrastive Learning: In response to problem 3 mentioned above,this paper adopts the contrastive learning method for research.This method uses a Siamese network to calculate the similarity between the features of different enhanced views of the same sample,which can perceive the feature differences between samples.The features learned by the Siamese network can well reflect the characteristics of the same class of samples and distinguish different categories of samples.Due to the network’s lack of ability to extract fine-grained features,a multi-scale module is used to improve the feature extraction process and obtain more accurate fine-grained features.The result shows that the classification model based on contrastive learning achieved an accuracy rate of 0.933 in the benign and malignant breast pathological image classification.This paper focuses on the use of deep learning for benign and malignant classification research in breast pathology images.To address the current issues of small sample sizes,insufficient use of spatial information,and low feature differences,classification methods based on multi-stage transfer learning,graph convolutional networks,and contrastive learning are proposed.The research results show that the above methods have achieved good performance in the benign and malignant classification of breast pathology images and have positive clinical application value.